EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Fix] Fix wrong backend chosen in hybrid backend (#8989)

Browse Source
This commit is contained in:
DarkSharpness
2025-08-08 21:21:17 -07:00
committed by GitHub
parent 7c0db868a1
commit fc42ff7b63
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/model_executor/model_runner.py
View File

@@ -378,6 +378,12 @@ class ModelRunner:
) )
server_args.attention_backend = "torch_native" server_args.attention_backend = "torch_native"
if server_args.prefill_attention_backend is not None and (
server_args.prefill_attention_backend
== server_args.decode_attention_backend
): # override the default attention backend
server_args.attention_backend = server_args.prefill_attention_backend
if server_args.attention_backend is None: if server_args.attention_backend is None:
""" """
Auto select the fastest attention backend. Auto select the fastest attention backend.